A friend in Indonesia recently told me about a conversation he had with ChatGPT. He had written a question in Bahasa Indonesia about how to handle a difficult family dispute. The chatbot responded fluently, in perfect Indonesian, with advice on communication strategies and conflict resolution. The grammar was impeccable. The tone was appropriate. And yet something was felt.
What HE offered was advice rooted in American cultural assumptions: prioritize your preferences, communicate directly and, if family members do not respect your boundaries, consider cutting them.
The answer was in Indonesian, but shaped by values that centered individual autonomy on it consensus buildingsocial harmony and collective family dynamics tend to be more important in Indonesian social life.
My friend was skeptical enough to notice the discrepancy and mention it to me. Many users may not. This is what prompted my research, published in International Review of Modern Sociologyin a pattern I found across major AI systems: Even when they were fluent in several languages, the language models retained their Western worldview. I call this “epistemological persistence.”
Fluency is not the same as comprehension
I have studied Indonesian societymedia and culture for over 30 years. This gives me a special vantage point for a problem that goes far beyond Indonesia: Large language models – LLM – including ChatGPT, Claude and Gemini can now speak dozens of languages with incredible fluency. This fluency creates the impression that AI understands local cultures.
However, the correct grammatical production of Indonesian, Arabic, Swahili or Hindi does not change the underlying worldview through which these systems reason. It doesn’t change how they think about people, relationships, responsibility, or what counts as a good outcome.
These assumptions are formed from training data derived primarily from English-language sources based in the United States. Meta’s open weight model LLaMA 2 was trained on about 89.7% of the English language text; LLaMA 3 includes only about 5% non-English data. The major commercial models do not publish equivalent divisions, but rely heavily on the same resources.
Arabic, the fifth most spoken language in the world, makes up less than 1% of the content on large training data sets. Languages with tens of millions of speakers, including Bengali and Hausa, barely appear.
Beneath the surface of these multilingual conversations, English functions as a hidden mediator. A study by researchers at the University of Oxford found that LLMs routinely conduct their core reasoning in Englisheven when requested in other languages. They translate the result in the final stage. A user gets perfect text in their preferred language, but the underlying logic originates elsewhere.
What the data shows
To examine how this happens in practice, I ran experiments with ChatGPT, Claude and Gemini. I asked questions in English and Indonesian about concepts such as education, responsibility, well-being, and some Indonesian terms that resist direct English translation. These included terms such as “mutual cooperation”, which describes a tradition of communal mutual aid.
Then I asked questions about education in both languages, using the word education in Indonesian. Responses consistently focused on individual development, personal autonomy, critical thinking and preparation for the labor market.
What mostly disappeared were dimensions of education that Indonesian educational traditions have historically emphasized. In Indonesia, education has long been focused on ethical discipline. Indonesian education scholars such as Christopher Bjork AND Robert Hefner HAVE documented how distinct these traditions are from models that mainly deal with education as a path to individual progress and career preparation, which is the lens through which AI tools viewed education.
The Indonesian concept of Embarrassed provides a more pointed example. Often translated as “shame” or “shame”, malu has been analyzed by anthropologists Clifford Geertz AND Tom Boellstorff like something closer to a common social awareness.
A person may feel bad when he speaks out of turn in front of elders or when a family member’s behavior reflects badly on the family. It regulates behavior and signals awareness of one’s position within a network of relationships.
It is cultivated, not just felt. It is a form of relational consciousness rather than a private psychological event.
When asked directly to define malu, the models acknowledged its social dimensions. However, in scenario-based questions that simply used the word without asking for a definition, all three reverted to the English translation of shame, consistently framing it as an individual emotional experience.
One representative response described malu as a normal emotional reaction that needs to be managed through self-reflection and confidence building – a personal psychological problem rather than a social one. The relational dimensions of the concept disappeared completely, being replaced by the language of individual emotional regulation.
A distinctly American worldview travels within the translation, largely unannounced.
Translation is much cheaper: Train a model on the wide web in English, then use multilingual production capabilities to serve global markets. As a media scholar Safiya Umoja Noble argues for algorithmic systems more broadly, what looks like a technical outcome is actually a structural outcome, shaped by who has the wealth and infrastructure to build these systems.
The embedded worldview is not a mistake; it is what happens when the production of knowledge seeks profit.
The main exceptions are Chinese models like DeepSeek and Qwen of Alibaba. However, they represent a real alternative to the US-dominated pipeline research shows they operate through a distinctly Chinese cultural lens. Asked about a workplace dispute, for example, they tend to advise silence or indirect phrases to maintain harmony rather than the direct private correction that Western models recommend.
Other regional efforts, such as THE SEA LION for Southeast Asia and Can-LLaMA for Hindi Kannada, use American patterns as a base. They add additional vocabulary and cultural information about local languages. But the core logic remains tied to the original US training.
Why this matters more than it might seem
One might reasonably ask if this is simply a limitation that users can work around. Decades of media studies demonstrate how audiences interpret foreign media through their cultural frames.
For example, anthropologist Brian Larkin documented how viewers rework in northern Nigeria narratives of Bollywood films to connect with local Islamic values. Larkin found that Muslim viewers in Kano reinterpreted Bollywood films through an Islamic moral lens, reading their narratives as reinforcing local values of propriety and ethical behavior. This dynamic depends on confronting the media as something of visible origin. But to do that, you need to know where your media is coming from.
Chat AI is different. Research at Harvard Business School reveals that people increasingly use AI systems for emotional support, advice and companionship. When a culturally specific worldview is delivered through a relationship that feels considerate and empathetic, in your language, it comes across less as a claim to be valued and more as a shared premise within a dialogue. It becomes hard to notice, and harder to contest.
The concern is that these perspectives will become the new normal. Some ways of reasoning about family life, education, and responsibility can become natural and self-evident. Language diversity among AI systems is real and growing. Cultural diversity of worldview, however, has not kept pace.
Gareth Barkin is a professor of anthropology and Asian studies, University of Puget Sound.
This article was reprinted from Conversation under a Creative Commons license. Read on original article.
